Claude 4 API: Tool Use Streaming & Prompt Caching vs Together AI Inference Endpoints

“The primitive here is clean: incremental tool-call deltas over SSE, and a cache-control header you attach to prompt segments to pin them server-side. The DX bet is that complexity lives in the HTTP layer, not in a new SDK abstraction — you opt in per-request, no new mental model required. The moment of truth is calling `stream=true` on a tool-use request and watching partial JSON arguments arrive before the model finishes thinking, which actually matters for agent loops where you want to dispatch work early. This is not a weekend-script replacement — implementing correct incremental JSON parsing for partial tool arguments plus a reliable distributed cache with 2M token capacity is a real engineering problem Anthropic has solved for you. The specific decision that earns the ship: cache invalidation is explicit and cache hits are reflected in the usage object, so you can actually measure what you're saving instead of guessing.”

“The primitive here is straightforward: dedicated compute allocation for open-source model inference with a contractual latency floor — not shared, not burstable, not 'best effort.' The DX bet is that production teams want to stop babysitting p99 latency graphs and just get a number they can put in their SLA doc. That's the right call. The moment of truth is when you point your production traffic at a dedicated endpoint and your tail latencies actually hold — and unlike shared inference pools, dedicated allocation means you're not racing your neighbors for GPU cycles. The weekend alternative (spinning your own vLLM on a reserved A100 instance) is absolutely real, but the SLA contract and the managed ops overhead is what you're paying for here. I'd want to see the actual SLA remediation terms before fully committing, but the core infrastructure bet is sound.”

“Direct competitors are OpenAI's cached completions and Google's context caching in Gemini 1.5 — both shipping for months — so Anthropic is catching up, not leading. The specific scenario where this breaks: cache hit rates depend entirely on prompt structure, and developers who dynamically compose system prompts (inserting user-specific context at the top) will see near-zero cache utilization and pay full price while assuming they're saving money. The prediction: this feature doesn't get killed — it becomes table stakes infrastructure and Anthropic wins by having the largest cache window (2M vs. competitors' current limits). What would have to be true for me to be wrong: OpenAI ships a 10M token cache window before Anthropic's ecosystem matures, commoditizing the advantage. Still a ship because the streaming tool-use delta is genuinely differentiated — no competitor has clean partial-argument streaming for tool calls yet, and that changes agent loop architecture in ways that matter.”

“Direct competitors are AWS Bedrock reserved throughput, Azure AI model deployments, and Fireworks AI — all of whom have been selling dedicated inference with latency guarantees for months. The specific scenario where Together breaks down is enterprise procurement: 'contact sales' pricing on the SLA tier means zero self-serve for the teams who need this most, and procurement cycles kill momentum. What kills this in 12 months is not a competitor — it's Llama 4 and Mistral becoming first-class citizens on hyperscaler managed services, at which point Together's open-source model advantage shrinks to a thin margin play. What earns the ship is that sub-100ms as a *contractual* commitment, not a marketing claim, is genuinely differentiated right now — if the remediation terms have teeth, this is real infrastructure.”

“The thesis this bets on: by 2027, the dominant AI application architecture is a persistent agent with a large, stable context (tools, memory, instructions) that gets reused across thousands of user interactions — making context I/O cost the primary unit economics lever, not generation cost. The dependency that has to hold: agents don't collapse back to stateless chatbots, and context windows keep growing faster than per-token prices fall. The second-order effect nobody's talking about: prompt caching at 2M tokens makes it economically viable to give every enterprise user a fully-loaded, role-specific agent context at request time — which shifts competitive differentiation from 'who has the best model' to 'who has the best cached context corpus,' effectively making knowledge curation the new moat. This tool is riding the trend of context-window expansion-as-infrastructure, and it's on-time, not early — but the streaming tool-use primitive is ahead of the curve on agent loop efficiency. The future state where this is infrastructure: every production agentic system has a cache manifest the same way it has a CDN config.”

“The thesis here is falsifiable: in 2-3 years, production AI applications will be built predominantly on open-source models, and the infrastructure layer that wins will be the one that offers hyperscaler-grade reliability guarantees without hyperscaler lock-in. For that to pay off, open-source model quality has to keep closing the gap with closed frontier models — which it's doing — and enterprises have to accept that running on third-party managed infrastructure for open-source is preferable to self-hosting, which is less certain. The second-order effect that matters: if contractual SLAs normalize for open-source inference, it removes the last credible objection enterprises have to not using GPT-4 or Claude — the 'we need guaranteed uptime and a contract' objection disappears. Together is on-time to this trend, not early, which means execution is everything and first-mover advantage is already gone.”

“The buyer is the engineering team at any company running Claude in production with long system prompts or multi-step agents — this comes out of the AI infrastructure budget, not a new budget line, which means no procurement friction. The pricing architecture is sound: cache reads at ~90% discount means the savings are real and measurable in the first billing cycle, which creates immediate retention — developers who restructure prompts to maximize cache hits are now architecturally coupled to Anthropic's caching implementation. The moat question is the honest one: this is infrastructure that OpenAI and Google will match, so the defensible position isn't the feature itself but the ecosystem of developers who've restructured their codebases around it. What survives a 10x model price drop: the streaming tool-use architecture, because that's about latency, not cost. The specific business decision that makes this viable is pricing cache reads as a separate SKU — it lets Anthropic capture value from high-volume production workloads without losing price-sensitive experimenters.”

“The buyer is clear — it's the ML infrastructure lead at a Series B+ company running open-source models in production — but the pricing architecture is not. 'Contact sales' for SLA tiers means Together is pricing this as an enterprise deal when the natural motion of developer-led AI tooling is self-serve with expansion. The moat question is real: Together's defensibility here is operational expertise running open-source models at scale, but that's a people moat, not a product moat. The moment Llama 4 gets native optimized inference on any hyperscaler with an SLA, Together has to compete on price alone. The business survives if they use dedicated endpoints as a wedge into enterprise contracts with broader platform consumption — but I don't see evidence that's the strategy, and a single product with contact-sales pricing is a services business dressed as a SaaS.”